We know deploying machine learning models can be tough. Today's models are notoriously bad at understanding their own risks -- they are biased on underrepresented data, brittle on challenging out-of-distribution scenarios, and can fail without warning when insufficiently trained.
Ensuring awareness of not one, but all of these risks, requires a tedious process involving changes to your model, its architecture, loss function, optimization procedure, and more.
Luckily, capsa has got you covered! Capsa automatically wraps your model (i.e., like a capsule!) and makes all of the internal changes so it can be end-to-end risk-aware. Capsa abstracts away all of those changes so you don't have to modify any of your existing training or deployment pipelines in order to build state-of-the-art trustworthy machine learning solutions.
capsa is available to be downloaded with Pip:
pip install capsaEager to make your models risk-aware? Let's go through a quick example of wrapping your model (e.g., using an MVEWrapper) to estimate risk from noise in your labels (i.e., aleatoric uncertainty).
import capsa
import tensorflow as tf
# Build your model
model = tf.keras.Sequential(...)
# Wrap the model with capsa to make it risk-aware.
# Capsa takes care of all the architecture, loss,
# and deployment changes so you don't have to!
model = capsa.MVEWrapper(model)
# Compile and train the wrapped model the
# same as you would have done with the
# original model. No changes!
model.compile(...)
model.fit(train_x, train_y, epochs=5)
# The model now outputs `RiskTensor` objects, which
# behave just like a normal `Tensor`, except they also
# contain multiple different quantitative risk measures.
pred_y = model(test_x)
# Returns the aleatoric uncertainty of this prediction
risk = pred_y.aleatoricHungry for more?
Checkout our tutorials on some more advanced functions with capsa including other forms of risk, composing wrappers together, high-dimensional datasets, and more! All tutorials can be opened directly in Google Collab so you can play around without needing access to GPUs.
Capsa is being actively maintained and advanced. It has been built with research, extensibility, and community development as a priority. We greatly appreciate contributions to the capsa repository and codebase, including issues, enhancements, and pull requests.
For more details please see here.
Capsa currently supports Keras Sequential models, although we are looking at possible solutions for supporting wider range of models. We can't wait to share you some new details soon!